xref: /linux/drivers/cpufreq/intel_pstate.c (revision b8d312aa075f33282565467662c4628dae0a2aff)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pstate.c: Native P state management for Intel processors
4  *
5  * (C) Copyright 2012 Intel Corporation
6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <trace/events/power.h>
28 
29 #include <asm/div64.h>
30 #include <asm/msr.h>
31 #include <asm/cpu_device_id.h>
32 #include <asm/cpufeature.h>
33 #include <asm/intel-family.h>
34 
35 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
36 
37 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
38 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
39 
40 #ifdef CONFIG_ACPI
41 #include <acpi/processor.h>
42 #include <acpi/cppc_acpi.h>
43 #endif
44 
45 #define FRAC_BITS 8
46 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
47 #define fp_toint(X) ((X) >> FRAC_BITS)
48 
49 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
50 
51 #define EXT_BITS 6
52 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
53 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
54 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
55 
56 static inline int32_t mul_fp(int32_t x, int32_t y)
57 {
58 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
59 }
60 
61 static inline int32_t div_fp(s64 x, s64 y)
62 {
63 	return div64_s64((int64_t)x << FRAC_BITS, y);
64 }
65 
66 static inline int ceiling_fp(int32_t x)
67 {
68 	int mask, ret;
69 
70 	ret = fp_toint(x);
71 	mask = (1 << FRAC_BITS) - 1;
72 	if (x & mask)
73 		ret += 1;
74 	return ret;
75 }
76 
77 static inline int32_t percent_fp(int percent)
78 {
79 	return div_fp(percent, 100);
80 }
81 
82 static inline u64 mul_ext_fp(u64 x, u64 y)
83 {
84 	return (x * y) >> EXT_FRAC_BITS;
85 }
86 
87 static inline u64 div_ext_fp(u64 x, u64 y)
88 {
89 	return div64_u64(x << EXT_FRAC_BITS, y);
90 }
91 
92 static inline int32_t percent_ext_fp(int percent)
93 {
94 	return div_ext_fp(percent, 100);
95 }
96 
97 /**
98  * struct sample -	Store performance sample
99  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
100  *			performance during last sample period
101  * @busy_scaled:	Scaled busy value which is used to calculate next
102  *			P state. This can be different than core_avg_perf
103  *			to account for cpu idle period
104  * @aperf:		Difference of actual performance frequency clock count
105  *			read from APERF MSR between last and current sample
106  * @mperf:		Difference of maximum performance frequency clock count
107  *			read from MPERF MSR between last and current sample
108  * @tsc:		Difference of time stamp counter between last and
109  *			current sample
110  * @time:		Current time from scheduler
111  *
112  * This structure is used in the cpudata structure to store performance sample
113  * data for choosing next P State.
114  */
115 struct sample {
116 	int32_t core_avg_perf;
117 	int32_t busy_scaled;
118 	u64 aperf;
119 	u64 mperf;
120 	u64 tsc;
121 	u64 time;
122 };
123 
124 /**
125  * struct pstate_data - Store P state data
126  * @current_pstate:	Current requested P state
127  * @min_pstate:		Min P state possible for this platform
128  * @max_pstate:		Max P state possible for this platform
129  * @max_pstate_physical:This is physical Max P state for a processor
130  *			This can be higher than the max_pstate which can
131  *			be limited by platform thermal design power limits
132  * @scaling:		Scaling factor to  convert frequency to cpufreq
133  *			frequency units
134  * @turbo_pstate:	Max Turbo P state possible for this platform
135  * @max_freq:		@max_pstate frequency in cpufreq units
136  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
137  *
138  * Stores the per cpu model P state limits and current P state.
139  */
140 struct pstate_data {
141 	int	current_pstate;
142 	int	min_pstate;
143 	int	max_pstate;
144 	int	max_pstate_physical;
145 	int	scaling;
146 	int	turbo_pstate;
147 	unsigned int max_freq;
148 	unsigned int turbo_freq;
149 };
150 
151 /**
152  * struct vid_data -	Stores voltage information data
153  * @min:		VID data for this platform corresponding to
154  *			the lowest P state
155  * @max:		VID data corresponding to the highest P State.
156  * @turbo:		VID data for turbo P state
157  * @ratio:		Ratio of (vid max - vid min) /
158  *			(max P state - Min P State)
159  *
160  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
161  * This data is used in Atom platforms, where in addition to target P state,
162  * the voltage data needs to be specified to select next P State.
163  */
164 struct vid_data {
165 	int min;
166 	int max;
167 	int turbo;
168 	int32_t ratio;
169 };
170 
171 /**
172  * struct global_params - Global parameters, mostly tunable via sysfs.
173  * @no_turbo:		Whether or not to use turbo P-states.
174  * @turbo_disabled:	Whethet or not turbo P-states are available at all,
175  *			based on the MSR_IA32_MISC_ENABLE value and whether or
176  *			not the maximum reported turbo P-state is different from
177  *			the maximum reported non-turbo one.
178  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
179  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
180  *			P-state capacity.
181  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
182  *			P-state capacity.
183  */
184 struct global_params {
185 	bool no_turbo;
186 	bool turbo_disabled;
187 	bool turbo_disabled_mf;
188 	int max_perf_pct;
189 	int min_perf_pct;
190 };
191 
192 /**
193  * struct cpudata -	Per CPU instance data storage
194  * @cpu:		CPU number for this instance data
195  * @policy:		CPUFreq policy value
196  * @update_util:	CPUFreq utility callback information
197  * @update_util_set:	CPUFreq utility callback is set
198  * @iowait_boost:	iowait-related boost fraction
199  * @last_update:	Time of the last update.
200  * @pstate:		Stores P state limits for this CPU
201  * @vid:		Stores VID limits for this CPU
202  * @last_sample_time:	Last Sample time
203  * @aperf_mperf_shift:	Number of clock cycles after aperf, merf is incremented
204  *			This shift is a multiplier to mperf delta to
205  *			calculate CPU busy.
206  * @prev_aperf:		Last APERF value read from APERF MSR
207  * @prev_mperf:		Last MPERF value read from MPERF MSR
208  * @prev_tsc:		Last timestamp counter (TSC) value
209  * @prev_cummulative_iowait: IO Wait time difference from last and
210  *			current sample
211  * @sample:		Storage for storing last Sample data
212  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
213  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
214  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
215  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
216  * @epp_powersave:	Last saved HWP energy performance preference
217  *			(EPP) or energy performance bias (EPB),
218  *			when policy switched to performance
219  * @epp_policy:		Last saved policy used to set EPP/EPB
220  * @epp_default:	Power on default HWP energy performance
221  *			preference/bias
222  * @epp_saved:		Saved EPP/EPB during system suspend or CPU offline
223  *			operation
224  * @hwp_req_cached:	Cached value of the last HWP Request MSR
225  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
226  * @last_io_update:	Last time when IO wake flag was set
227  * @sched_flags:	Store scheduler flags for possible cross CPU update
228  * @hwp_boost_min:	Last HWP boosted min performance
229  *
230  * This structure stores per CPU instance data for all CPUs.
231  */
232 struct cpudata {
233 	int cpu;
234 
235 	unsigned int policy;
236 	struct update_util_data update_util;
237 	bool   update_util_set;
238 
239 	struct pstate_data pstate;
240 	struct vid_data vid;
241 
242 	u64	last_update;
243 	u64	last_sample_time;
244 	u64	aperf_mperf_shift;
245 	u64	prev_aperf;
246 	u64	prev_mperf;
247 	u64	prev_tsc;
248 	u64	prev_cummulative_iowait;
249 	struct sample sample;
250 	int32_t	min_perf_ratio;
251 	int32_t	max_perf_ratio;
252 #ifdef CONFIG_ACPI
253 	struct acpi_processor_performance acpi_perf_data;
254 	bool valid_pss_table;
255 #endif
256 	unsigned int iowait_boost;
257 	s16 epp_powersave;
258 	s16 epp_policy;
259 	s16 epp_default;
260 	s16 epp_saved;
261 	u64 hwp_req_cached;
262 	u64 hwp_cap_cached;
263 	u64 last_io_update;
264 	unsigned int sched_flags;
265 	u32 hwp_boost_min;
266 };
267 
268 static struct cpudata **all_cpu_data;
269 
270 /**
271  * struct pstate_funcs - Per CPU model specific callbacks
272  * @get_max:		Callback to get maximum non turbo effective P state
273  * @get_max_physical:	Callback to get maximum non turbo physical P state
274  * @get_min:		Callback to get minimum P state
275  * @get_turbo:		Callback to get turbo P state
276  * @get_scaling:	Callback to get frequency scaling factor
277  * @get_val:		Callback to convert P state to actual MSR write value
278  * @get_vid:		Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284 	int (*get_max)(void);
285 	int (*get_max_physical)(void);
286 	int (*get_min)(void);
287 	int (*get_turbo)(void);
288 	int (*get_scaling)(void);
289 	int (*get_aperf_mperf_shift)(void);
290 	u64 (*get_val)(struct cpudata*, int pstate);
291 	void (*get_vid)(struct cpudata *);
292 };
293 
294 static struct pstate_funcs pstate_funcs __read_mostly;
295 
296 static int hwp_active __read_mostly;
297 static int hwp_mode_bdw __read_mostly;
298 static bool per_cpu_limits __read_mostly;
299 static bool hwp_boost __read_mostly;
300 
301 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
302 
303 #ifdef CONFIG_ACPI
304 static bool acpi_ppc;
305 #endif
306 
307 static struct global_params global;
308 
309 static DEFINE_MUTEX(intel_pstate_driver_lock);
310 static DEFINE_MUTEX(intel_pstate_limits_lock);
311 
312 #ifdef CONFIG_ACPI
313 
314 static bool intel_pstate_acpi_pm_profile_server(void)
315 {
316 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
317 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
318 		return true;
319 
320 	return false;
321 }
322 
323 static bool intel_pstate_get_ppc_enable_status(void)
324 {
325 	if (intel_pstate_acpi_pm_profile_server())
326 		return true;
327 
328 	return acpi_ppc;
329 }
330 
331 #ifdef CONFIG_ACPI_CPPC_LIB
332 
333 /* The work item is needed to avoid CPU hotplug locking issues */
334 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
335 {
336 	sched_set_itmt_support();
337 }
338 
339 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
340 
341 static void intel_pstate_set_itmt_prio(int cpu)
342 {
343 	struct cppc_perf_caps cppc_perf;
344 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
345 	int ret;
346 
347 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
348 	if (ret)
349 		return;
350 
351 	/*
352 	 * The priorities can be set regardless of whether or not
353 	 * sched_set_itmt_support(true) has been called and it is valid to
354 	 * update them at any time after it has been called.
355 	 */
356 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
357 
358 	if (max_highest_perf <= min_highest_perf) {
359 		if (cppc_perf.highest_perf > max_highest_perf)
360 			max_highest_perf = cppc_perf.highest_perf;
361 
362 		if (cppc_perf.highest_perf < min_highest_perf)
363 			min_highest_perf = cppc_perf.highest_perf;
364 
365 		if (max_highest_perf > min_highest_perf) {
366 			/*
367 			 * This code can be run during CPU online under the
368 			 * CPU hotplug locks, so sched_set_itmt_support()
369 			 * cannot be called from here.  Queue up a work item
370 			 * to invoke it.
371 			 */
372 			schedule_work(&sched_itmt_work);
373 		}
374 	}
375 }
376 
377 static int intel_pstate_get_cppc_guranteed(int cpu)
378 {
379 	struct cppc_perf_caps cppc_perf;
380 	int ret;
381 
382 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
383 	if (ret)
384 		return ret;
385 
386 	if (cppc_perf.guaranteed_perf)
387 		return cppc_perf.guaranteed_perf;
388 
389 	return cppc_perf.nominal_perf;
390 }
391 
392 #else /* CONFIG_ACPI_CPPC_LIB */
393 static void intel_pstate_set_itmt_prio(int cpu)
394 {
395 }
396 #endif /* CONFIG_ACPI_CPPC_LIB */
397 
398 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
399 {
400 	struct cpudata *cpu;
401 	int ret;
402 	int i;
403 
404 	if (hwp_active) {
405 		intel_pstate_set_itmt_prio(policy->cpu);
406 		return;
407 	}
408 
409 	if (!intel_pstate_get_ppc_enable_status())
410 		return;
411 
412 	cpu = all_cpu_data[policy->cpu];
413 
414 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
415 						  policy->cpu);
416 	if (ret)
417 		return;
418 
419 	/*
420 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
421 	 * guarantee that the states returned by it map to the states in our
422 	 * list directly.
423 	 */
424 	if (cpu->acpi_perf_data.control_register.space_id !=
425 						ACPI_ADR_SPACE_FIXED_HARDWARE)
426 		goto err;
427 
428 	/*
429 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
430 	 * usual without taking _PSS into account
431 	 */
432 	if (cpu->acpi_perf_data.state_count < 2)
433 		goto err;
434 
435 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
436 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
437 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
438 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
439 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
440 			 (u32) cpu->acpi_perf_data.states[i].power,
441 			 (u32) cpu->acpi_perf_data.states[i].control);
442 	}
443 
444 	/*
445 	 * The _PSS table doesn't contain whole turbo frequency range.
446 	 * This just contains +1 MHZ above the max non turbo frequency,
447 	 * with control value corresponding to max turbo ratio. But
448 	 * when cpufreq set policy is called, it will call with this
449 	 * max frequency, which will cause a reduced performance as
450 	 * this driver uses real max turbo frequency as the max
451 	 * frequency. So correct this frequency in _PSS table to
452 	 * correct max turbo frequency based on the turbo state.
453 	 * Also need to convert to MHz as _PSS freq is in MHz.
454 	 */
455 	if (!global.turbo_disabled)
456 		cpu->acpi_perf_data.states[0].core_frequency =
457 					policy->cpuinfo.max_freq / 1000;
458 	cpu->valid_pss_table = true;
459 	pr_debug("_PPC limits will be enforced\n");
460 
461 	return;
462 
463  err:
464 	cpu->valid_pss_table = false;
465 	acpi_processor_unregister_performance(policy->cpu);
466 }
467 
468 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
469 {
470 	struct cpudata *cpu;
471 
472 	cpu = all_cpu_data[policy->cpu];
473 	if (!cpu->valid_pss_table)
474 		return;
475 
476 	acpi_processor_unregister_performance(policy->cpu);
477 }
478 #else /* CONFIG_ACPI */
479 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
480 {
481 }
482 
483 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
484 {
485 }
486 
487 static inline bool intel_pstate_acpi_pm_profile_server(void)
488 {
489 	return false;
490 }
491 #endif /* CONFIG_ACPI */
492 
493 #ifndef CONFIG_ACPI_CPPC_LIB
494 static int intel_pstate_get_cppc_guranteed(int cpu)
495 {
496 	return -ENOTSUPP;
497 }
498 #endif /* CONFIG_ACPI_CPPC_LIB */
499 
500 static inline void update_turbo_state(void)
501 {
502 	u64 misc_en;
503 	struct cpudata *cpu;
504 
505 	cpu = all_cpu_data[0];
506 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
507 	global.turbo_disabled =
508 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
509 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
510 }
511 
512 static int min_perf_pct_min(void)
513 {
514 	struct cpudata *cpu = all_cpu_data[0];
515 	int turbo_pstate = cpu->pstate.turbo_pstate;
516 
517 	return turbo_pstate ?
518 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
519 }
520 
521 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
522 {
523 	u64 epb;
524 	int ret;
525 
526 	if (!boot_cpu_has(X86_FEATURE_EPB))
527 		return -ENXIO;
528 
529 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
530 	if (ret)
531 		return (s16)ret;
532 
533 	return (s16)(epb & 0x0f);
534 }
535 
536 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
537 {
538 	s16 epp;
539 
540 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
541 		/*
542 		 * When hwp_req_data is 0, means that caller didn't read
543 		 * MSR_HWP_REQUEST, so need to read and get EPP.
544 		 */
545 		if (!hwp_req_data) {
546 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
547 					    &hwp_req_data);
548 			if (epp)
549 				return epp;
550 		}
551 		epp = (hwp_req_data >> 24) & 0xff;
552 	} else {
553 		/* When there is no EPP present, HWP uses EPB settings */
554 		epp = intel_pstate_get_epb(cpu_data);
555 	}
556 
557 	return epp;
558 }
559 
560 static int intel_pstate_set_epb(int cpu, s16 pref)
561 {
562 	u64 epb;
563 	int ret;
564 
565 	if (!boot_cpu_has(X86_FEATURE_EPB))
566 		return -ENXIO;
567 
568 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
569 	if (ret)
570 		return ret;
571 
572 	epb = (epb & ~0x0f) | pref;
573 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
574 
575 	return 0;
576 }
577 
578 /*
579  * EPP/EPB display strings corresponding to EPP index in the
580  * energy_perf_strings[]
581  *	index		String
582  *-------------------------------------
583  *	0		default
584  *	1		performance
585  *	2		balance_performance
586  *	3		balance_power
587  *	4		power
588  */
589 static const char * const energy_perf_strings[] = {
590 	"default",
591 	"performance",
592 	"balance_performance",
593 	"balance_power",
594 	"power",
595 	NULL
596 };
597 static const unsigned int epp_values[] = {
598 	HWP_EPP_PERFORMANCE,
599 	HWP_EPP_BALANCE_PERFORMANCE,
600 	HWP_EPP_BALANCE_POWERSAVE,
601 	HWP_EPP_POWERSAVE
602 };
603 
604 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
605 {
606 	s16 epp;
607 	int index = -EINVAL;
608 
609 	epp = intel_pstate_get_epp(cpu_data, 0);
610 	if (epp < 0)
611 		return epp;
612 
613 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
614 		if (epp == HWP_EPP_PERFORMANCE)
615 			return 1;
616 		if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
617 			return 2;
618 		if (epp <= HWP_EPP_BALANCE_POWERSAVE)
619 			return 3;
620 		else
621 			return 4;
622 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
623 		/*
624 		 * Range:
625 		 *	0x00-0x03	:	Performance
626 		 *	0x04-0x07	:	Balance performance
627 		 *	0x08-0x0B	:	Balance power
628 		 *	0x0C-0x0F	:	Power
629 		 * The EPB is a 4 bit value, but our ranges restrict the
630 		 * value which can be set. Here only using top two bits
631 		 * effectively.
632 		 */
633 		index = (epp >> 2) + 1;
634 	}
635 
636 	return index;
637 }
638 
639 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
640 					      int pref_index)
641 {
642 	int epp = -EINVAL;
643 	int ret;
644 
645 	if (!pref_index)
646 		epp = cpu_data->epp_default;
647 
648 	mutex_lock(&intel_pstate_limits_lock);
649 
650 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
651 		u64 value;
652 
653 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
654 		if (ret)
655 			goto return_pref;
656 
657 		value &= ~GENMASK_ULL(31, 24);
658 
659 		if (epp == -EINVAL)
660 			epp = epp_values[pref_index - 1];
661 
662 		value |= (u64)epp << 24;
663 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
664 	} else {
665 		if (epp == -EINVAL)
666 			epp = (pref_index - 1) << 2;
667 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
668 	}
669 return_pref:
670 	mutex_unlock(&intel_pstate_limits_lock);
671 
672 	return ret;
673 }
674 
675 static ssize_t show_energy_performance_available_preferences(
676 				struct cpufreq_policy *policy, char *buf)
677 {
678 	int i = 0;
679 	int ret = 0;
680 
681 	while (energy_perf_strings[i] != NULL)
682 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
683 
684 	ret += sprintf(&buf[ret], "\n");
685 
686 	return ret;
687 }
688 
689 cpufreq_freq_attr_ro(energy_performance_available_preferences);
690 
691 static ssize_t store_energy_performance_preference(
692 		struct cpufreq_policy *policy, const char *buf, size_t count)
693 {
694 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
695 	char str_preference[21];
696 	int ret;
697 
698 	ret = sscanf(buf, "%20s", str_preference);
699 	if (ret != 1)
700 		return -EINVAL;
701 
702 	ret = match_string(energy_perf_strings, -1, str_preference);
703 	if (ret < 0)
704 		return ret;
705 
706 	intel_pstate_set_energy_pref_index(cpu_data, ret);
707 	return count;
708 }
709 
710 static ssize_t show_energy_performance_preference(
711 				struct cpufreq_policy *policy, char *buf)
712 {
713 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
714 	int preference;
715 
716 	preference = intel_pstate_get_energy_pref_index(cpu_data);
717 	if (preference < 0)
718 		return preference;
719 
720 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
721 }
722 
723 cpufreq_freq_attr_rw(energy_performance_preference);
724 
725 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
726 {
727 	struct cpudata *cpu;
728 	u64 cap;
729 	int ratio;
730 
731 	ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
732 	if (ratio <= 0) {
733 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
734 		ratio = HWP_GUARANTEED_PERF(cap);
735 	}
736 
737 	cpu = all_cpu_data[policy->cpu];
738 
739 	return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
740 }
741 
742 cpufreq_freq_attr_ro(base_frequency);
743 
744 static struct freq_attr *hwp_cpufreq_attrs[] = {
745 	&energy_performance_preference,
746 	&energy_performance_available_preferences,
747 	&base_frequency,
748 	NULL,
749 };
750 
751 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
752 				     int *current_max)
753 {
754 	u64 cap;
755 
756 	rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
757 	WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
758 	if (global.no_turbo)
759 		*current_max = HWP_GUARANTEED_PERF(cap);
760 	else
761 		*current_max = HWP_HIGHEST_PERF(cap);
762 
763 	*phy_max = HWP_HIGHEST_PERF(cap);
764 }
765 
766 static void intel_pstate_hwp_set(unsigned int cpu)
767 {
768 	struct cpudata *cpu_data = all_cpu_data[cpu];
769 	int max, min;
770 	u64 value;
771 	s16 epp;
772 
773 	max = cpu_data->max_perf_ratio;
774 	min = cpu_data->min_perf_ratio;
775 
776 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
777 		min = max;
778 
779 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
780 
781 	value &= ~HWP_MIN_PERF(~0L);
782 	value |= HWP_MIN_PERF(min);
783 
784 	value &= ~HWP_MAX_PERF(~0L);
785 	value |= HWP_MAX_PERF(max);
786 
787 	if (cpu_data->epp_policy == cpu_data->policy)
788 		goto skip_epp;
789 
790 	cpu_data->epp_policy = cpu_data->policy;
791 
792 	if (cpu_data->epp_saved >= 0) {
793 		epp = cpu_data->epp_saved;
794 		cpu_data->epp_saved = -EINVAL;
795 		goto update_epp;
796 	}
797 
798 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
799 		epp = intel_pstate_get_epp(cpu_data, value);
800 		cpu_data->epp_powersave = epp;
801 		/* If EPP read was failed, then don't try to write */
802 		if (epp < 0)
803 			goto skip_epp;
804 
805 		epp = 0;
806 	} else {
807 		/* skip setting EPP, when saved value is invalid */
808 		if (cpu_data->epp_powersave < 0)
809 			goto skip_epp;
810 
811 		/*
812 		 * No need to restore EPP when it is not zero. This
813 		 * means:
814 		 *  - Policy is not changed
815 		 *  - user has manually changed
816 		 *  - Error reading EPB
817 		 */
818 		epp = intel_pstate_get_epp(cpu_data, value);
819 		if (epp)
820 			goto skip_epp;
821 
822 		epp = cpu_data->epp_powersave;
823 	}
824 update_epp:
825 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
826 		value &= ~GENMASK_ULL(31, 24);
827 		value |= (u64)epp << 24;
828 	} else {
829 		intel_pstate_set_epb(cpu, epp);
830 	}
831 skip_epp:
832 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
833 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
834 }
835 
836 static void intel_pstate_hwp_force_min_perf(int cpu)
837 {
838 	u64 value;
839 	int min_perf;
840 
841 	value = all_cpu_data[cpu]->hwp_req_cached;
842 	value &= ~GENMASK_ULL(31, 0);
843 	min_perf = HWP_LOWEST_PERF(all_cpu_data[cpu]->hwp_cap_cached);
844 
845 	/* Set hwp_max = hwp_min */
846 	value |= HWP_MAX_PERF(min_perf);
847 	value |= HWP_MIN_PERF(min_perf);
848 
849 	/* Set EPP/EPB to min */
850 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
851 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
852 	else
853 		intel_pstate_set_epb(cpu, HWP_EPP_BALANCE_POWERSAVE);
854 
855 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
856 }
857 
858 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
859 {
860 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
861 
862 	if (!hwp_active)
863 		return 0;
864 
865 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
866 
867 	return 0;
868 }
869 
870 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
871 
872 static int intel_pstate_resume(struct cpufreq_policy *policy)
873 {
874 	if (!hwp_active)
875 		return 0;
876 
877 	mutex_lock(&intel_pstate_limits_lock);
878 
879 	if (policy->cpu == 0)
880 		intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
881 
882 	all_cpu_data[policy->cpu]->epp_policy = 0;
883 	intel_pstate_hwp_set(policy->cpu);
884 
885 	mutex_unlock(&intel_pstate_limits_lock);
886 
887 	return 0;
888 }
889 
890 static void intel_pstate_update_policies(void)
891 {
892 	int cpu;
893 
894 	for_each_possible_cpu(cpu)
895 		cpufreq_update_policy(cpu);
896 }
897 
898 static void intel_pstate_update_max_freq(unsigned int cpu)
899 {
900 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
901 	struct cpudata *cpudata;
902 
903 	if (!policy)
904 		return;
905 
906 	cpudata = all_cpu_data[cpu];
907 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
908 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
909 
910 	refresh_frequency_limits(policy);
911 
912 	cpufreq_cpu_release(policy);
913 }
914 
915 static void intel_pstate_update_limits(unsigned int cpu)
916 {
917 	mutex_lock(&intel_pstate_driver_lock);
918 
919 	update_turbo_state();
920 	/*
921 	 * If turbo has been turned on or off globally, policy limits for
922 	 * all CPUs need to be updated to reflect that.
923 	 */
924 	if (global.turbo_disabled_mf != global.turbo_disabled) {
925 		global.turbo_disabled_mf = global.turbo_disabled;
926 		for_each_possible_cpu(cpu)
927 			intel_pstate_update_max_freq(cpu);
928 	} else {
929 		cpufreq_update_policy(cpu);
930 	}
931 
932 	mutex_unlock(&intel_pstate_driver_lock);
933 }
934 
935 /************************** sysfs begin ************************/
936 #define show_one(file_name, object)					\
937 	static ssize_t show_##file_name					\
938 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
939 	{								\
940 		return sprintf(buf, "%u\n", global.object);		\
941 	}
942 
943 static ssize_t intel_pstate_show_status(char *buf);
944 static int intel_pstate_update_status(const char *buf, size_t size);
945 
946 static ssize_t show_status(struct kobject *kobj,
947 			   struct kobj_attribute *attr, char *buf)
948 {
949 	ssize_t ret;
950 
951 	mutex_lock(&intel_pstate_driver_lock);
952 	ret = intel_pstate_show_status(buf);
953 	mutex_unlock(&intel_pstate_driver_lock);
954 
955 	return ret;
956 }
957 
958 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
959 			    const char *buf, size_t count)
960 {
961 	char *p = memchr(buf, '\n', count);
962 	int ret;
963 
964 	mutex_lock(&intel_pstate_driver_lock);
965 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
966 	mutex_unlock(&intel_pstate_driver_lock);
967 
968 	return ret < 0 ? ret : count;
969 }
970 
971 static ssize_t show_turbo_pct(struct kobject *kobj,
972 				struct kobj_attribute *attr, char *buf)
973 {
974 	struct cpudata *cpu;
975 	int total, no_turbo, turbo_pct;
976 	uint32_t turbo_fp;
977 
978 	mutex_lock(&intel_pstate_driver_lock);
979 
980 	if (!intel_pstate_driver) {
981 		mutex_unlock(&intel_pstate_driver_lock);
982 		return -EAGAIN;
983 	}
984 
985 	cpu = all_cpu_data[0];
986 
987 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
988 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
989 	turbo_fp = div_fp(no_turbo, total);
990 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
991 
992 	mutex_unlock(&intel_pstate_driver_lock);
993 
994 	return sprintf(buf, "%u\n", turbo_pct);
995 }
996 
997 static ssize_t show_num_pstates(struct kobject *kobj,
998 				struct kobj_attribute *attr, char *buf)
999 {
1000 	struct cpudata *cpu;
1001 	int total;
1002 
1003 	mutex_lock(&intel_pstate_driver_lock);
1004 
1005 	if (!intel_pstate_driver) {
1006 		mutex_unlock(&intel_pstate_driver_lock);
1007 		return -EAGAIN;
1008 	}
1009 
1010 	cpu = all_cpu_data[0];
1011 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1012 
1013 	mutex_unlock(&intel_pstate_driver_lock);
1014 
1015 	return sprintf(buf, "%u\n", total);
1016 }
1017 
1018 static ssize_t show_no_turbo(struct kobject *kobj,
1019 			     struct kobj_attribute *attr, char *buf)
1020 {
1021 	ssize_t ret;
1022 
1023 	mutex_lock(&intel_pstate_driver_lock);
1024 
1025 	if (!intel_pstate_driver) {
1026 		mutex_unlock(&intel_pstate_driver_lock);
1027 		return -EAGAIN;
1028 	}
1029 
1030 	update_turbo_state();
1031 	if (global.turbo_disabled)
1032 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1033 	else
1034 		ret = sprintf(buf, "%u\n", global.no_turbo);
1035 
1036 	mutex_unlock(&intel_pstate_driver_lock);
1037 
1038 	return ret;
1039 }
1040 
1041 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1042 			      const char *buf, size_t count)
1043 {
1044 	unsigned int input;
1045 	int ret;
1046 
1047 	ret = sscanf(buf, "%u", &input);
1048 	if (ret != 1)
1049 		return -EINVAL;
1050 
1051 	mutex_lock(&intel_pstate_driver_lock);
1052 
1053 	if (!intel_pstate_driver) {
1054 		mutex_unlock(&intel_pstate_driver_lock);
1055 		return -EAGAIN;
1056 	}
1057 
1058 	mutex_lock(&intel_pstate_limits_lock);
1059 
1060 	update_turbo_state();
1061 	if (global.turbo_disabled) {
1062 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1063 		mutex_unlock(&intel_pstate_limits_lock);
1064 		mutex_unlock(&intel_pstate_driver_lock);
1065 		return -EPERM;
1066 	}
1067 
1068 	global.no_turbo = clamp_t(int, input, 0, 1);
1069 
1070 	if (global.no_turbo) {
1071 		struct cpudata *cpu = all_cpu_data[0];
1072 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1073 
1074 		/* Squash the global minimum into the permitted range. */
1075 		if (global.min_perf_pct > pct)
1076 			global.min_perf_pct = pct;
1077 	}
1078 
1079 	mutex_unlock(&intel_pstate_limits_lock);
1080 
1081 	intel_pstate_update_policies();
1082 
1083 	mutex_unlock(&intel_pstate_driver_lock);
1084 
1085 	return count;
1086 }
1087 
1088 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1089 				  const char *buf, size_t count)
1090 {
1091 	unsigned int input;
1092 	int ret;
1093 
1094 	ret = sscanf(buf, "%u", &input);
1095 	if (ret != 1)
1096 		return -EINVAL;
1097 
1098 	mutex_lock(&intel_pstate_driver_lock);
1099 
1100 	if (!intel_pstate_driver) {
1101 		mutex_unlock(&intel_pstate_driver_lock);
1102 		return -EAGAIN;
1103 	}
1104 
1105 	mutex_lock(&intel_pstate_limits_lock);
1106 
1107 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1108 
1109 	mutex_unlock(&intel_pstate_limits_lock);
1110 
1111 	intel_pstate_update_policies();
1112 
1113 	mutex_unlock(&intel_pstate_driver_lock);
1114 
1115 	return count;
1116 }
1117 
1118 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1119 				  const char *buf, size_t count)
1120 {
1121 	unsigned int input;
1122 	int ret;
1123 
1124 	ret = sscanf(buf, "%u", &input);
1125 	if (ret != 1)
1126 		return -EINVAL;
1127 
1128 	mutex_lock(&intel_pstate_driver_lock);
1129 
1130 	if (!intel_pstate_driver) {
1131 		mutex_unlock(&intel_pstate_driver_lock);
1132 		return -EAGAIN;
1133 	}
1134 
1135 	mutex_lock(&intel_pstate_limits_lock);
1136 
1137 	global.min_perf_pct = clamp_t(int, input,
1138 				      min_perf_pct_min(), global.max_perf_pct);
1139 
1140 	mutex_unlock(&intel_pstate_limits_lock);
1141 
1142 	intel_pstate_update_policies();
1143 
1144 	mutex_unlock(&intel_pstate_driver_lock);
1145 
1146 	return count;
1147 }
1148 
1149 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1150 				struct kobj_attribute *attr, char *buf)
1151 {
1152 	return sprintf(buf, "%u\n", hwp_boost);
1153 }
1154 
1155 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1156 				       struct kobj_attribute *b,
1157 				       const char *buf, size_t count)
1158 {
1159 	unsigned int input;
1160 	int ret;
1161 
1162 	ret = kstrtouint(buf, 10, &input);
1163 	if (ret)
1164 		return ret;
1165 
1166 	mutex_lock(&intel_pstate_driver_lock);
1167 	hwp_boost = !!input;
1168 	intel_pstate_update_policies();
1169 	mutex_unlock(&intel_pstate_driver_lock);
1170 
1171 	return count;
1172 }
1173 
1174 show_one(max_perf_pct, max_perf_pct);
1175 show_one(min_perf_pct, min_perf_pct);
1176 
1177 define_one_global_rw(status);
1178 define_one_global_rw(no_turbo);
1179 define_one_global_rw(max_perf_pct);
1180 define_one_global_rw(min_perf_pct);
1181 define_one_global_ro(turbo_pct);
1182 define_one_global_ro(num_pstates);
1183 define_one_global_rw(hwp_dynamic_boost);
1184 
1185 static struct attribute *intel_pstate_attributes[] = {
1186 	&status.attr,
1187 	&no_turbo.attr,
1188 	&turbo_pct.attr,
1189 	&num_pstates.attr,
1190 	NULL
1191 };
1192 
1193 static const struct attribute_group intel_pstate_attr_group = {
1194 	.attrs = intel_pstate_attributes,
1195 };
1196 
1197 static void __init intel_pstate_sysfs_expose_params(void)
1198 {
1199 	struct kobject *intel_pstate_kobject;
1200 	int rc;
1201 
1202 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1203 						&cpu_subsys.dev_root->kobj);
1204 	if (WARN_ON(!intel_pstate_kobject))
1205 		return;
1206 
1207 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1208 	if (WARN_ON(rc))
1209 		return;
1210 
1211 	/*
1212 	 * If per cpu limits are enforced there are no global limits, so
1213 	 * return without creating max/min_perf_pct attributes
1214 	 */
1215 	if (per_cpu_limits)
1216 		return;
1217 
1218 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1219 	WARN_ON(rc);
1220 
1221 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1222 	WARN_ON(rc);
1223 
1224 	if (hwp_active) {
1225 		rc = sysfs_create_file(intel_pstate_kobject,
1226 				       &hwp_dynamic_boost.attr);
1227 		WARN_ON(rc);
1228 	}
1229 }
1230 /************************** sysfs end ************************/
1231 
1232 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1233 {
1234 	/* First disable HWP notification interrupt as we don't process them */
1235 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1236 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1237 
1238 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1239 	cpudata->epp_policy = 0;
1240 	if (cpudata->epp_default == -EINVAL)
1241 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1242 }
1243 
1244 #define MSR_IA32_POWER_CTL_BIT_EE	19
1245 
1246 /* Disable energy efficiency optimization */
1247 static void intel_pstate_disable_ee(int cpu)
1248 {
1249 	u64 power_ctl;
1250 	int ret;
1251 
1252 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1253 	if (ret)
1254 		return;
1255 
1256 	if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1257 		pr_info("Disabling energy efficiency optimization\n");
1258 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1259 		wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1260 	}
1261 }
1262 
1263 static int atom_get_min_pstate(void)
1264 {
1265 	u64 value;
1266 
1267 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1268 	return (value >> 8) & 0x7F;
1269 }
1270 
1271 static int atom_get_max_pstate(void)
1272 {
1273 	u64 value;
1274 
1275 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1276 	return (value >> 16) & 0x7F;
1277 }
1278 
1279 static int atom_get_turbo_pstate(void)
1280 {
1281 	u64 value;
1282 
1283 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1284 	return value & 0x7F;
1285 }
1286 
1287 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1288 {
1289 	u64 val;
1290 	int32_t vid_fp;
1291 	u32 vid;
1292 
1293 	val = (u64)pstate << 8;
1294 	if (global.no_turbo && !global.turbo_disabled)
1295 		val |= (u64)1 << 32;
1296 
1297 	vid_fp = cpudata->vid.min + mul_fp(
1298 		int_tofp(pstate - cpudata->pstate.min_pstate),
1299 		cpudata->vid.ratio);
1300 
1301 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1302 	vid = ceiling_fp(vid_fp);
1303 
1304 	if (pstate > cpudata->pstate.max_pstate)
1305 		vid = cpudata->vid.turbo;
1306 
1307 	return val | vid;
1308 }
1309 
1310 static int silvermont_get_scaling(void)
1311 {
1312 	u64 value;
1313 	int i;
1314 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1315 	static int silvermont_freq_table[] = {
1316 		83300, 100000, 133300, 116700, 80000};
1317 
1318 	rdmsrl(MSR_FSB_FREQ, value);
1319 	i = value & 0x7;
1320 	WARN_ON(i > 4);
1321 
1322 	return silvermont_freq_table[i];
1323 }
1324 
1325 static int airmont_get_scaling(void)
1326 {
1327 	u64 value;
1328 	int i;
1329 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1330 	static int airmont_freq_table[] = {
1331 		83300, 100000, 133300, 116700, 80000,
1332 		93300, 90000, 88900, 87500};
1333 
1334 	rdmsrl(MSR_FSB_FREQ, value);
1335 	i = value & 0xF;
1336 	WARN_ON(i > 8);
1337 
1338 	return airmont_freq_table[i];
1339 }
1340 
1341 static void atom_get_vid(struct cpudata *cpudata)
1342 {
1343 	u64 value;
1344 
1345 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1346 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1347 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1348 	cpudata->vid.ratio = div_fp(
1349 		cpudata->vid.max - cpudata->vid.min,
1350 		int_tofp(cpudata->pstate.max_pstate -
1351 			cpudata->pstate.min_pstate));
1352 
1353 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1354 	cpudata->vid.turbo = value & 0x7f;
1355 }
1356 
1357 static int core_get_min_pstate(void)
1358 {
1359 	u64 value;
1360 
1361 	rdmsrl(MSR_PLATFORM_INFO, value);
1362 	return (value >> 40) & 0xFF;
1363 }
1364 
1365 static int core_get_max_pstate_physical(void)
1366 {
1367 	u64 value;
1368 
1369 	rdmsrl(MSR_PLATFORM_INFO, value);
1370 	return (value >> 8) & 0xFF;
1371 }
1372 
1373 static int core_get_tdp_ratio(u64 plat_info)
1374 {
1375 	/* Check how many TDP levels present */
1376 	if (plat_info & 0x600000000) {
1377 		u64 tdp_ctrl;
1378 		u64 tdp_ratio;
1379 		int tdp_msr;
1380 		int err;
1381 
1382 		/* Get the TDP level (0, 1, 2) to get ratios */
1383 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1384 		if (err)
1385 			return err;
1386 
1387 		/* TDP MSR are continuous starting at 0x648 */
1388 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1389 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1390 		if (err)
1391 			return err;
1392 
1393 		/* For level 1 and 2, bits[23:16] contain the ratio */
1394 		if (tdp_ctrl & 0x03)
1395 			tdp_ratio >>= 16;
1396 
1397 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1398 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1399 
1400 		return (int)tdp_ratio;
1401 	}
1402 
1403 	return -ENXIO;
1404 }
1405 
1406 static int core_get_max_pstate(void)
1407 {
1408 	u64 tar;
1409 	u64 plat_info;
1410 	int max_pstate;
1411 	int tdp_ratio;
1412 	int err;
1413 
1414 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1415 	max_pstate = (plat_info >> 8) & 0xFF;
1416 
1417 	tdp_ratio = core_get_tdp_ratio(plat_info);
1418 	if (tdp_ratio <= 0)
1419 		return max_pstate;
1420 
1421 	if (hwp_active) {
1422 		/* Turbo activation ratio is not used on HWP platforms */
1423 		return tdp_ratio;
1424 	}
1425 
1426 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1427 	if (!err) {
1428 		int tar_levels;
1429 
1430 		/* Do some sanity checking for safety */
1431 		tar_levels = tar & 0xff;
1432 		if (tdp_ratio - 1 == tar_levels) {
1433 			max_pstate = tar_levels;
1434 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1435 		}
1436 	}
1437 
1438 	return max_pstate;
1439 }
1440 
1441 static int core_get_turbo_pstate(void)
1442 {
1443 	u64 value;
1444 	int nont, ret;
1445 
1446 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1447 	nont = core_get_max_pstate();
1448 	ret = (value) & 255;
1449 	if (ret <= nont)
1450 		ret = nont;
1451 	return ret;
1452 }
1453 
1454 static inline int core_get_scaling(void)
1455 {
1456 	return 100000;
1457 }
1458 
1459 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1460 {
1461 	u64 val;
1462 
1463 	val = (u64)pstate << 8;
1464 	if (global.no_turbo && !global.turbo_disabled)
1465 		val |= (u64)1 << 32;
1466 
1467 	return val;
1468 }
1469 
1470 static int knl_get_aperf_mperf_shift(void)
1471 {
1472 	return 10;
1473 }
1474 
1475 static int knl_get_turbo_pstate(void)
1476 {
1477 	u64 value;
1478 	int nont, ret;
1479 
1480 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1481 	nont = core_get_max_pstate();
1482 	ret = (((value) >> 8) & 0xFF);
1483 	if (ret <= nont)
1484 		ret = nont;
1485 	return ret;
1486 }
1487 
1488 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1489 {
1490 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1491 	cpu->pstate.current_pstate = pstate;
1492 	/*
1493 	 * Generally, there is no guarantee that this code will always run on
1494 	 * the CPU being updated, so force the register update to run on the
1495 	 * right CPU.
1496 	 */
1497 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1498 		      pstate_funcs.get_val(cpu, pstate));
1499 }
1500 
1501 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1502 {
1503 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1504 }
1505 
1506 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1507 {
1508 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1509 
1510 	update_turbo_state();
1511 	intel_pstate_set_pstate(cpu, pstate);
1512 }
1513 
1514 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1515 {
1516 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1517 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1518 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1519 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1520 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1521 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1522 
1523 	if (hwp_active && !hwp_mode_bdw) {
1524 		unsigned int phy_max, current_max;
1525 
1526 		intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1527 		cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1528 	} else {
1529 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1530 	}
1531 
1532 	if (pstate_funcs.get_aperf_mperf_shift)
1533 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1534 
1535 	if (pstate_funcs.get_vid)
1536 		pstate_funcs.get_vid(cpu);
1537 
1538 	intel_pstate_set_min_pstate(cpu);
1539 }
1540 
1541 /*
1542  * Long hold time will keep high perf limits for long time,
1543  * which negatively impacts perf/watt for some workloads,
1544  * like specpower. 3ms is based on experiements on some
1545  * workoads.
1546  */
1547 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1548 
1549 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1550 {
1551 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1552 	u32 max_limit = (hwp_req & 0xff00) >> 8;
1553 	u32 min_limit = (hwp_req & 0xff);
1554 	u32 boost_level1;
1555 
1556 	/*
1557 	 * Cases to consider (User changes via sysfs or boot time):
1558 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1559 	 *	No boost, return.
1560 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1561 	 *     Should result in one level boost only for P0.
1562 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1563 	 *     Should result in two level boost:
1564 	 *         (min + p1)/2 and P1.
1565 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1566 	 *     Should result in three level boost:
1567 	 *        (min + p1)/2, P1 and P0.
1568 	 */
1569 
1570 	/* If max and min are equal or already at max, nothing to boost */
1571 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1572 		return;
1573 
1574 	if (!cpu->hwp_boost_min)
1575 		cpu->hwp_boost_min = min_limit;
1576 
1577 	/* level at half way mark between min and guranteed */
1578 	boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1579 
1580 	if (cpu->hwp_boost_min < boost_level1)
1581 		cpu->hwp_boost_min = boost_level1;
1582 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1583 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1584 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1585 		 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1586 		cpu->hwp_boost_min = max_limit;
1587 	else
1588 		return;
1589 
1590 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1591 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
1592 	cpu->last_update = cpu->sample.time;
1593 }
1594 
1595 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1596 {
1597 	if (cpu->hwp_boost_min) {
1598 		bool expired;
1599 
1600 		/* Check if we are idle for hold time to boost down */
1601 		expired = time_after64(cpu->sample.time, cpu->last_update +
1602 				       hwp_boost_hold_time_ns);
1603 		if (expired) {
1604 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1605 			cpu->hwp_boost_min = 0;
1606 		}
1607 	}
1608 	cpu->last_update = cpu->sample.time;
1609 }
1610 
1611 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1612 						      u64 time)
1613 {
1614 	cpu->sample.time = time;
1615 
1616 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1617 		bool do_io = false;
1618 
1619 		cpu->sched_flags = 0;
1620 		/*
1621 		 * Set iowait_boost flag and update time. Since IO WAIT flag
1622 		 * is set all the time, we can't just conclude that there is
1623 		 * some IO bound activity is scheduled on this CPU with just
1624 		 * one occurrence. If we receive at least two in two
1625 		 * consecutive ticks, then we treat as boost candidate.
1626 		 */
1627 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1628 			do_io = true;
1629 
1630 		cpu->last_io_update = time;
1631 
1632 		if (do_io)
1633 			intel_pstate_hwp_boost_up(cpu);
1634 
1635 	} else {
1636 		intel_pstate_hwp_boost_down(cpu);
1637 	}
1638 }
1639 
1640 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1641 						u64 time, unsigned int flags)
1642 {
1643 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1644 
1645 	cpu->sched_flags |= flags;
1646 
1647 	if (smp_processor_id() == cpu->cpu)
1648 		intel_pstate_update_util_hwp_local(cpu, time);
1649 }
1650 
1651 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1652 {
1653 	struct sample *sample = &cpu->sample;
1654 
1655 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1656 }
1657 
1658 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1659 {
1660 	u64 aperf, mperf;
1661 	unsigned long flags;
1662 	u64 tsc;
1663 
1664 	local_irq_save(flags);
1665 	rdmsrl(MSR_IA32_APERF, aperf);
1666 	rdmsrl(MSR_IA32_MPERF, mperf);
1667 	tsc = rdtsc();
1668 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1669 		local_irq_restore(flags);
1670 		return false;
1671 	}
1672 	local_irq_restore(flags);
1673 
1674 	cpu->last_sample_time = cpu->sample.time;
1675 	cpu->sample.time = time;
1676 	cpu->sample.aperf = aperf;
1677 	cpu->sample.mperf = mperf;
1678 	cpu->sample.tsc =  tsc;
1679 	cpu->sample.aperf -= cpu->prev_aperf;
1680 	cpu->sample.mperf -= cpu->prev_mperf;
1681 	cpu->sample.tsc -= cpu->prev_tsc;
1682 
1683 	cpu->prev_aperf = aperf;
1684 	cpu->prev_mperf = mperf;
1685 	cpu->prev_tsc = tsc;
1686 	/*
1687 	 * First time this function is invoked in a given cycle, all of the
1688 	 * previous sample data fields are equal to zero or stale and they must
1689 	 * be populated with meaningful numbers for things to work, so assume
1690 	 * that sample.time will always be reset before setting the utilization
1691 	 * update hook and make the caller skip the sample then.
1692 	 */
1693 	if (cpu->last_sample_time) {
1694 		intel_pstate_calc_avg_perf(cpu);
1695 		return true;
1696 	}
1697 	return false;
1698 }
1699 
1700 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1701 {
1702 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1703 }
1704 
1705 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1706 {
1707 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1708 			  cpu->sample.core_avg_perf);
1709 }
1710 
1711 static inline int32_t get_target_pstate(struct cpudata *cpu)
1712 {
1713 	struct sample *sample = &cpu->sample;
1714 	int32_t busy_frac;
1715 	int target, avg_pstate;
1716 
1717 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1718 			   sample->tsc);
1719 
1720 	if (busy_frac < cpu->iowait_boost)
1721 		busy_frac = cpu->iowait_boost;
1722 
1723 	sample->busy_scaled = busy_frac * 100;
1724 
1725 	target = global.no_turbo || global.turbo_disabled ?
1726 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1727 	target += target >> 2;
1728 	target = mul_fp(target, busy_frac);
1729 	if (target < cpu->pstate.min_pstate)
1730 		target = cpu->pstate.min_pstate;
1731 
1732 	/*
1733 	 * If the average P-state during the previous cycle was higher than the
1734 	 * current target, add 50% of the difference to the target to reduce
1735 	 * possible performance oscillations and offset possible performance
1736 	 * loss related to moving the workload from one CPU to another within
1737 	 * a package/module.
1738 	 */
1739 	avg_pstate = get_avg_pstate(cpu);
1740 	if (avg_pstate > target)
1741 		target += (avg_pstate - target) >> 1;
1742 
1743 	return target;
1744 }
1745 
1746 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1747 {
1748 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1749 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1750 
1751 	return clamp_t(int, pstate, min_pstate, max_pstate);
1752 }
1753 
1754 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1755 {
1756 	if (pstate == cpu->pstate.current_pstate)
1757 		return;
1758 
1759 	cpu->pstate.current_pstate = pstate;
1760 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1761 }
1762 
1763 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1764 {
1765 	int from = cpu->pstate.current_pstate;
1766 	struct sample *sample;
1767 	int target_pstate;
1768 
1769 	update_turbo_state();
1770 
1771 	target_pstate = get_target_pstate(cpu);
1772 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1773 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1774 	intel_pstate_update_pstate(cpu, target_pstate);
1775 
1776 	sample = &cpu->sample;
1777 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1778 		fp_toint(sample->busy_scaled),
1779 		from,
1780 		cpu->pstate.current_pstate,
1781 		sample->mperf,
1782 		sample->aperf,
1783 		sample->tsc,
1784 		get_avg_frequency(cpu),
1785 		fp_toint(cpu->iowait_boost * 100));
1786 }
1787 
1788 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1789 				     unsigned int flags)
1790 {
1791 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1792 	u64 delta_ns;
1793 
1794 	/* Don't allow remote callbacks */
1795 	if (smp_processor_id() != cpu->cpu)
1796 		return;
1797 
1798 	delta_ns = time - cpu->last_update;
1799 	if (flags & SCHED_CPUFREQ_IOWAIT) {
1800 		/* Start over if the CPU may have been idle. */
1801 		if (delta_ns > TICK_NSEC) {
1802 			cpu->iowait_boost = ONE_EIGHTH_FP;
1803 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
1804 			cpu->iowait_boost <<= 1;
1805 			if (cpu->iowait_boost > int_tofp(1))
1806 				cpu->iowait_boost = int_tofp(1);
1807 		} else {
1808 			cpu->iowait_boost = ONE_EIGHTH_FP;
1809 		}
1810 	} else if (cpu->iowait_boost) {
1811 		/* Clear iowait_boost if the CPU may have been idle. */
1812 		if (delta_ns > TICK_NSEC)
1813 			cpu->iowait_boost = 0;
1814 		else
1815 			cpu->iowait_boost >>= 1;
1816 	}
1817 	cpu->last_update = time;
1818 	delta_ns = time - cpu->sample.time;
1819 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1820 		return;
1821 
1822 	if (intel_pstate_sample(cpu, time))
1823 		intel_pstate_adjust_pstate(cpu);
1824 }
1825 
1826 static struct pstate_funcs core_funcs = {
1827 	.get_max = core_get_max_pstate,
1828 	.get_max_physical = core_get_max_pstate_physical,
1829 	.get_min = core_get_min_pstate,
1830 	.get_turbo = core_get_turbo_pstate,
1831 	.get_scaling = core_get_scaling,
1832 	.get_val = core_get_val,
1833 };
1834 
1835 static const struct pstate_funcs silvermont_funcs = {
1836 	.get_max = atom_get_max_pstate,
1837 	.get_max_physical = atom_get_max_pstate,
1838 	.get_min = atom_get_min_pstate,
1839 	.get_turbo = atom_get_turbo_pstate,
1840 	.get_val = atom_get_val,
1841 	.get_scaling = silvermont_get_scaling,
1842 	.get_vid = atom_get_vid,
1843 };
1844 
1845 static const struct pstate_funcs airmont_funcs = {
1846 	.get_max = atom_get_max_pstate,
1847 	.get_max_physical = atom_get_max_pstate,
1848 	.get_min = atom_get_min_pstate,
1849 	.get_turbo = atom_get_turbo_pstate,
1850 	.get_val = atom_get_val,
1851 	.get_scaling = airmont_get_scaling,
1852 	.get_vid = atom_get_vid,
1853 };
1854 
1855 static const struct pstate_funcs knl_funcs = {
1856 	.get_max = core_get_max_pstate,
1857 	.get_max_physical = core_get_max_pstate_physical,
1858 	.get_min = core_get_min_pstate,
1859 	.get_turbo = knl_get_turbo_pstate,
1860 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1861 	.get_scaling = core_get_scaling,
1862 	.get_val = core_get_val,
1863 };
1864 
1865 #define ICPU(model, policy) \
1866 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1867 			(unsigned long)&policy }
1868 
1869 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1870 	ICPU(INTEL_FAM6_SANDYBRIDGE, 		core_funcs),
1871 	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		core_funcs),
1872 	ICPU(INTEL_FAM6_ATOM_SILVERMONT,	silvermont_funcs),
1873 	ICPU(INTEL_FAM6_IVYBRIDGE,		core_funcs),
1874 	ICPU(INTEL_FAM6_HASWELL_CORE,		core_funcs),
1875 	ICPU(INTEL_FAM6_BROADWELL_CORE,		core_funcs),
1876 	ICPU(INTEL_FAM6_IVYBRIDGE_X,		core_funcs),
1877 	ICPU(INTEL_FAM6_HASWELL_X,		core_funcs),
1878 	ICPU(INTEL_FAM6_HASWELL_ULT,		core_funcs),
1879 	ICPU(INTEL_FAM6_HASWELL_GT3E,		core_funcs),
1880 	ICPU(INTEL_FAM6_BROADWELL_GT3E,		core_funcs),
1881 	ICPU(INTEL_FAM6_ATOM_AIRMONT,		airmont_funcs),
1882 	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		core_funcs),
1883 	ICPU(INTEL_FAM6_BROADWELL_X,		core_funcs),
1884 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	core_funcs),
1885 	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	core_funcs),
1886 	ICPU(INTEL_FAM6_XEON_PHI_KNL,		knl_funcs),
1887 	ICPU(INTEL_FAM6_XEON_PHI_KNM,		knl_funcs),
1888 	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		core_funcs),
1889 	ICPU(INTEL_FAM6_ATOM_GOLDMONT_PLUS,     core_funcs),
1890 	ICPU(INTEL_FAM6_SKYLAKE_X,		core_funcs),
1891 	{}
1892 };
1893 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1894 
1895 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1896 	ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1897 	ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1898 	ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1899 	{}
1900 };
1901 
1902 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1903 	ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs),
1904 	{}
1905 };
1906 
1907 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1908 	ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1909 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
1910 	{}
1911 };
1912 
1913 static int intel_pstate_init_cpu(unsigned int cpunum)
1914 {
1915 	struct cpudata *cpu;
1916 
1917 	cpu = all_cpu_data[cpunum];
1918 
1919 	if (!cpu) {
1920 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1921 		if (!cpu)
1922 			return -ENOMEM;
1923 
1924 		all_cpu_data[cpunum] = cpu;
1925 
1926 		cpu->epp_default = -EINVAL;
1927 		cpu->epp_powersave = -EINVAL;
1928 		cpu->epp_saved = -EINVAL;
1929 	}
1930 
1931 	cpu = all_cpu_data[cpunum];
1932 
1933 	cpu->cpu = cpunum;
1934 
1935 	if (hwp_active) {
1936 		const struct x86_cpu_id *id;
1937 
1938 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1939 		if (id)
1940 			intel_pstate_disable_ee(cpunum);
1941 
1942 		intel_pstate_hwp_enable(cpu);
1943 
1944 		id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1945 		if (id && intel_pstate_acpi_pm_profile_server())
1946 			hwp_boost = true;
1947 	}
1948 
1949 	intel_pstate_get_cpu_pstates(cpu);
1950 
1951 	pr_debug("controlling: cpu %d\n", cpunum);
1952 
1953 	return 0;
1954 }
1955 
1956 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1957 {
1958 	struct cpudata *cpu = all_cpu_data[cpu_num];
1959 
1960 	if (hwp_active && !hwp_boost)
1961 		return;
1962 
1963 	if (cpu->update_util_set)
1964 		return;
1965 
1966 	/* Prevent intel_pstate_update_util() from using stale data. */
1967 	cpu->sample.time = 0;
1968 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1969 				     (hwp_active ?
1970 				      intel_pstate_update_util_hwp :
1971 				      intel_pstate_update_util));
1972 	cpu->update_util_set = true;
1973 }
1974 
1975 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1976 {
1977 	struct cpudata *cpu_data = all_cpu_data[cpu];
1978 
1979 	if (!cpu_data->update_util_set)
1980 		return;
1981 
1982 	cpufreq_remove_update_util_hook(cpu);
1983 	cpu_data->update_util_set = false;
1984 	synchronize_rcu();
1985 }
1986 
1987 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1988 {
1989 	return global.turbo_disabled || global.no_turbo ?
1990 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1991 }
1992 
1993 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1994 					    struct cpudata *cpu)
1995 {
1996 	int max_freq = intel_pstate_get_max_freq(cpu);
1997 	int32_t max_policy_perf, min_policy_perf;
1998 	int max_state, turbo_max;
1999 
2000 	/*
2001 	 * HWP needs some special consideration, because on BDX the
2002 	 * HWP_REQUEST uses abstract value to represent performance
2003 	 * rather than pure ratios.
2004 	 */
2005 	if (hwp_active) {
2006 		intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2007 	} else {
2008 		max_state = global.no_turbo || global.turbo_disabled ?
2009 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2010 		turbo_max = cpu->pstate.turbo_pstate;
2011 	}
2012 
2013 	max_policy_perf = max_state * policy->max / max_freq;
2014 	if (policy->max == policy->min) {
2015 		min_policy_perf = max_policy_perf;
2016 	} else {
2017 		min_policy_perf = max_state * policy->min / max_freq;
2018 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2019 					  0, max_policy_perf);
2020 	}
2021 
2022 	pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
2023 		 policy->cpu, max_state,
2024 		 min_policy_perf, max_policy_perf);
2025 
2026 	/* Normalize user input to [min_perf, max_perf] */
2027 	if (per_cpu_limits) {
2028 		cpu->min_perf_ratio = min_policy_perf;
2029 		cpu->max_perf_ratio = max_policy_perf;
2030 	} else {
2031 		int32_t global_min, global_max;
2032 
2033 		/* Global limits are in percent of the maximum turbo P-state. */
2034 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2035 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2036 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2037 
2038 		pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu,
2039 			 global_min, global_max);
2040 
2041 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2042 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2043 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2044 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2045 
2046 		/* Make sure min_perf <= max_perf */
2047 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2048 					  cpu->max_perf_ratio);
2049 
2050 	}
2051 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu,
2052 		 cpu->max_perf_ratio,
2053 		 cpu->min_perf_ratio);
2054 }
2055 
2056 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2057 {
2058 	struct cpudata *cpu;
2059 
2060 	if (!policy->cpuinfo.max_freq)
2061 		return -ENODEV;
2062 
2063 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2064 		 policy->cpuinfo.max_freq, policy->max);
2065 
2066 	cpu = all_cpu_data[policy->cpu];
2067 	cpu->policy = policy->policy;
2068 
2069 	mutex_lock(&intel_pstate_limits_lock);
2070 
2071 	intel_pstate_update_perf_limits(policy, cpu);
2072 
2073 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2074 		/*
2075 		 * NOHZ_FULL CPUs need this as the governor callback may not
2076 		 * be invoked on them.
2077 		 */
2078 		intel_pstate_clear_update_util_hook(policy->cpu);
2079 		intel_pstate_max_within_limits(cpu);
2080 	} else {
2081 		intel_pstate_set_update_util_hook(policy->cpu);
2082 	}
2083 
2084 	if (hwp_active) {
2085 		/*
2086 		 * When hwp_boost was active before and dynamically it
2087 		 * was turned off, in that case we need to clear the
2088 		 * update util hook.
2089 		 */
2090 		if (!hwp_boost)
2091 			intel_pstate_clear_update_util_hook(policy->cpu);
2092 		intel_pstate_hwp_set(policy->cpu);
2093 	}
2094 
2095 	mutex_unlock(&intel_pstate_limits_lock);
2096 
2097 	return 0;
2098 }
2099 
2100 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
2101 					 struct cpudata *cpu)
2102 {
2103 	if (!hwp_active &&
2104 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2105 	    policy->max < policy->cpuinfo.max_freq &&
2106 	    policy->max > cpu->pstate.max_freq) {
2107 		pr_debug("policy->max > max non turbo frequency\n");
2108 		policy->max = policy->cpuinfo.max_freq;
2109 	}
2110 }
2111 
2112 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2113 {
2114 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2115 
2116 	update_turbo_state();
2117 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2118 				     intel_pstate_get_max_freq(cpu));
2119 
2120 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2121 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2122 		return -EINVAL;
2123 
2124 	intel_pstate_adjust_policy_max(policy, cpu);
2125 
2126 	return 0;
2127 }
2128 
2129 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2130 {
2131 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2132 }
2133 
2134 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2135 {
2136 	pr_debug("CPU %d exiting\n", policy->cpu);
2137 
2138 	intel_pstate_clear_update_util_hook(policy->cpu);
2139 	if (hwp_active) {
2140 		intel_pstate_hwp_save_state(policy);
2141 		intel_pstate_hwp_force_min_perf(policy->cpu);
2142 	} else {
2143 		intel_cpufreq_stop_cpu(policy);
2144 	}
2145 }
2146 
2147 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2148 {
2149 	intel_pstate_exit_perf_limits(policy);
2150 
2151 	policy->fast_switch_possible = false;
2152 
2153 	return 0;
2154 }
2155 
2156 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2157 {
2158 	struct cpudata *cpu;
2159 	int rc;
2160 
2161 	rc = intel_pstate_init_cpu(policy->cpu);
2162 	if (rc)
2163 		return rc;
2164 
2165 	cpu = all_cpu_data[policy->cpu];
2166 
2167 	cpu->max_perf_ratio = 0xFF;
2168 	cpu->min_perf_ratio = 0;
2169 
2170 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2171 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2172 
2173 	/* cpuinfo and default policy values */
2174 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2175 	update_turbo_state();
2176 	global.turbo_disabled_mf = global.turbo_disabled;
2177 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2178 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2179 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2180 
2181 	if (hwp_active) {
2182 		unsigned int max_freq;
2183 
2184 		max_freq = global.turbo_disabled ?
2185 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2186 		if (max_freq < policy->cpuinfo.max_freq)
2187 			policy->cpuinfo.max_freq = max_freq;
2188 	}
2189 
2190 	intel_pstate_init_acpi_perf_limits(policy);
2191 
2192 	policy->fast_switch_possible = true;
2193 
2194 	return 0;
2195 }
2196 
2197 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2198 {
2199 	int ret = __intel_pstate_cpu_init(policy);
2200 
2201 	if (ret)
2202 		return ret;
2203 
2204 	if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2205 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2206 	else
2207 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
2208 
2209 	return 0;
2210 }
2211 
2212 static struct cpufreq_driver intel_pstate = {
2213 	.flags		= CPUFREQ_CONST_LOOPS,
2214 	.verify		= intel_pstate_verify_policy,
2215 	.setpolicy	= intel_pstate_set_policy,
2216 	.suspend	= intel_pstate_hwp_save_state,
2217 	.resume		= intel_pstate_resume,
2218 	.init		= intel_pstate_cpu_init,
2219 	.exit		= intel_pstate_cpu_exit,
2220 	.stop_cpu	= intel_pstate_stop_cpu,
2221 	.update_limits	= intel_pstate_update_limits,
2222 	.name		= "intel_pstate",
2223 };
2224 
2225 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2226 {
2227 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2228 
2229 	update_turbo_state();
2230 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2231 				     intel_pstate_get_max_freq(cpu));
2232 
2233 	intel_pstate_adjust_policy_max(policy, cpu);
2234 
2235 	intel_pstate_update_perf_limits(policy, cpu);
2236 
2237 	return 0;
2238 }
2239 
2240 /* Use of trace in passive mode:
2241  *
2242  * In passive mode the trace core_busy field (also known as the
2243  * performance field, and lablelled as such on the graphs; also known as
2244  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2245  * driver call was via the normal or fast switch path. Various graphs
2246  * output from the intel_pstate_tracer.py utility that include core_busy
2247  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2248  * so we use 10 to indicate the the normal path through the driver, and
2249  * 90 to indicate the fast switch path through the driver.
2250  * The scaled_busy field is not used, and is set to 0.
2251  */
2252 
2253 #define	INTEL_PSTATE_TRACE_TARGET 10
2254 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2255 
2256 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2257 {
2258 	struct sample *sample;
2259 
2260 	if (!trace_pstate_sample_enabled())
2261 		return;
2262 
2263 	if (!intel_pstate_sample(cpu, ktime_get()))
2264 		return;
2265 
2266 	sample = &cpu->sample;
2267 	trace_pstate_sample(trace_type,
2268 		0,
2269 		old_pstate,
2270 		cpu->pstate.current_pstate,
2271 		sample->mperf,
2272 		sample->aperf,
2273 		sample->tsc,
2274 		get_avg_frequency(cpu),
2275 		fp_toint(cpu->iowait_boost * 100));
2276 }
2277 
2278 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2279 				unsigned int target_freq,
2280 				unsigned int relation)
2281 {
2282 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2283 	struct cpufreq_freqs freqs;
2284 	int target_pstate, old_pstate;
2285 
2286 	update_turbo_state();
2287 
2288 	freqs.old = policy->cur;
2289 	freqs.new = target_freq;
2290 
2291 	cpufreq_freq_transition_begin(policy, &freqs);
2292 	switch (relation) {
2293 	case CPUFREQ_RELATION_L:
2294 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2295 		break;
2296 	case CPUFREQ_RELATION_H:
2297 		target_pstate = freqs.new / cpu->pstate.scaling;
2298 		break;
2299 	default:
2300 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2301 		break;
2302 	}
2303 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2304 	old_pstate = cpu->pstate.current_pstate;
2305 	if (target_pstate != cpu->pstate.current_pstate) {
2306 		cpu->pstate.current_pstate = target_pstate;
2307 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2308 			      pstate_funcs.get_val(cpu, target_pstate));
2309 	}
2310 	freqs.new = target_pstate * cpu->pstate.scaling;
2311 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2312 	cpufreq_freq_transition_end(policy, &freqs, false);
2313 
2314 	return 0;
2315 }
2316 
2317 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2318 					      unsigned int target_freq)
2319 {
2320 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2321 	int target_pstate, old_pstate;
2322 
2323 	update_turbo_state();
2324 
2325 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2326 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2327 	old_pstate = cpu->pstate.current_pstate;
2328 	intel_pstate_update_pstate(cpu, target_pstate);
2329 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2330 	return target_pstate * cpu->pstate.scaling;
2331 }
2332 
2333 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2334 {
2335 	int ret = __intel_pstate_cpu_init(policy);
2336 
2337 	if (ret)
2338 		return ret;
2339 
2340 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2341 	policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2342 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2343 	policy->cur = policy->cpuinfo.min_freq;
2344 
2345 	return 0;
2346 }
2347 
2348 static struct cpufreq_driver intel_cpufreq = {
2349 	.flags		= CPUFREQ_CONST_LOOPS,
2350 	.verify		= intel_cpufreq_verify_policy,
2351 	.target		= intel_cpufreq_target,
2352 	.fast_switch	= intel_cpufreq_fast_switch,
2353 	.init		= intel_cpufreq_cpu_init,
2354 	.exit		= intel_pstate_cpu_exit,
2355 	.stop_cpu	= intel_cpufreq_stop_cpu,
2356 	.update_limits	= intel_pstate_update_limits,
2357 	.name		= "intel_cpufreq",
2358 };
2359 
2360 static struct cpufreq_driver *default_driver = &intel_pstate;
2361 
2362 static void intel_pstate_driver_cleanup(void)
2363 {
2364 	unsigned int cpu;
2365 
2366 	get_online_cpus();
2367 	for_each_online_cpu(cpu) {
2368 		if (all_cpu_data[cpu]) {
2369 			if (intel_pstate_driver == &intel_pstate)
2370 				intel_pstate_clear_update_util_hook(cpu);
2371 
2372 			kfree(all_cpu_data[cpu]);
2373 			all_cpu_data[cpu] = NULL;
2374 		}
2375 	}
2376 	put_online_cpus();
2377 	intel_pstate_driver = NULL;
2378 }
2379 
2380 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2381 {
2382 	int ret;
2383 
2384 	memset(&global, 0, sizeof(global));
2385 	global.max_perf_pct = 100;
2386 
2387 	intel_pstate_driver = driver;
2388 	ret = cpufreq_register_driver(intel_pstate_driver);
2389 	if (ret) {
2390 		intel_pstate_driver_cleanup();
2391 		return ret;
2392 	}
2393 
2394 	global.min_perf_pct = min_perf_pct_min();
2395 
2396 	return 0;
2397 }
2398 
2399 static int intel_pstate_unregister_driver(void)
2400 {
2401 	if (hwp_active)
2402 		return -EBUSY;
2403 
2404 	cpufreq_unregister_driver(intel_pstate_driver);
2405 	intel_pstate_driver_cleanup();
2406 
2407 	return 0;
2408 }
2409 
2410 static ssize_t intel_pstate_show_status(char *buf)
2411 {
2412 	if (!intel_pstate_driver)
2413 		return sprintf(buf, "off\n");
2414 
2415 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2416 					"active" : "passive");
2417 }
2418 
2419 static int intel_pstate_update_status(const char *buf, size_t size)
2420 {
2421 	int ret;
2422 
2423 	if (size == 3 && !strncmp(buf, "off", size))
2424 		return intel_pstate_driver ?
2425 			intel_pstate_unregister_driver() : -EINVAL;
2426 
2427 	if (size == 6 && !strncmp(buf, "active", size)) {
2428 		if (intel_pstate_driver) {
2429 			if (intel_pstate_driver == &intel_pstate)
2430 				return 0;
2431 
2432 			ret = intel_pstate_unregister_driver();
2433 			if (ret)
2434 				return ret;
2435 		}
2436 
2437 		return intel_pstate_register_driver(&intel_pstate);
2438 	}
2439 
2440 	if (size == 7 && !strncmp(buf, "passive", size)) {
2441 		if (intel_pstate_driver) {
2442 			if (intel_pstate_driver == &intel_cpufreq)
2443 				return 0;
2444 
2445 			ret = intel_pstate_unregister_driver();
2446 			if (ret)
2447 				return ret;
2448 		}
2449 
2450 		return intel_pstate_register_driver(&intel_cpufreq);
2451 	}
2452 
2453 	return -EINVAL;
2454 }
2455 
2456 static int no_load __initdata;
2457 static int no_hwp __initdata;
2458 static int hwp_only __initdata;
2459 static unsigned int force_load __initdata;
2460 
2461 static int __init intel_pstate_msrs_not_valid(void)
2462 {
2463 	if (!pstate_funcs.get_max() ||
2464 	    !pstate_funcs.get_min() ||
2465 	    !pstate_funcs.get_turbo())
2466 		return -ENODEV;
2467 
2468 	return 0;
2469 }
2470 
2471 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2472 {
2473 	pstate_funcs.get_max   = funcs->get_max;
2474 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2475 	pstate_funcs.get_min   = funcs->get_min;
2476 	pstate_funcs.get_turbo = funcs->get_turbo;
2477 	pstate_funcs.get_scaling = funcs->get_scaling;
2478 	pstate_funcs.get_val   = funcs->get_val;
2479 	pstate_funcs.get_vid   = funcs->get_vid;
2480 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2481 }
2482 
2483 #ifdef CONFIG_ACPI
2484 
2485 static bool __init intel_pstate_no_acpi_pss(void)
2486 {
2487 	int i;
2488 
2489 	for_each_possible_cpu(i) {
2490 		acpi_status status;
2491 		union acpi_object *pss;
2492 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2493 		struct acpi_processor *pr = per_cpu(processors, i);
2494 
2495 		if (!pr)
2496 			continue;
2497 
2498 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2499 		if (ACPI_FAILURE(status))
2500 			continue;
2501 
2502 		pss = buffer.pointer;
2503 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2504 			kfree(pss);
2505 			return false;
2506 		}
2507 
2508 		kfree(pss);
2509 	}
2510 
2511 	pr_debug("ACPI _PSS not found\n");
2512 	return true;
2513 }
2514 
2515 static bool __init intel_pstate_no_acpi_pcch(void)
2516 {
2517 	acpi_status status;
2518 	acpi_handle handle;
2519 
2520 	status = acpi_get_handle(NULL, "\\_SB", &handle);
2521 	if (ACPI_FAILURE(status))
2522 		goto not_found;
2523 
2524 	if (acpi_has_method(handle, "PCCH"))
2525 		return false;
2526 
2527 not_found:
2528 	pr_debug("ACPI PCCH not found\n");
2529 	return true;
2530 }
2531 
2532 static bool __init intel_pstate_has_acpi_ppc(void)
2533 {
2534 	int i;
2535 
2536 	for_each_possible_cpu(i) {
2537 		struct acpi_processor *pr = per_cpu(processors, i);
2538 
2539 		if (!pr)
2540 			continue;
2541 		if (acpi_has_method(pr->handle, "_PPC"))
2542 			return true;
2543 	}
2544 	pr_debug("ACPI _PPC not found\n");
2545 	return false;
2546 }
2547 
2548 enum {
2549 	PSS,
2550 	PPC,
2551 };
2552 
2553 /* Hardware vendor-specific info that has its own power management modes */
2554 static struct acpi_platform_list plat_info[] __initdata = {
2555 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, 0, PSS},
2556 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2557 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2558 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2559 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2560 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2561 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2562 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2563 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2564 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2565 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2566 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2567 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2568 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2569 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2570 	{ } /* End */
2571 };
2572 
2573 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2574 {
2575 	const struct x86_cpu_id *id;
2576 	u64 misc_pwr;
2577 	int idx;
2578 
2579 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2580 	if (id) {
2581 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2582 		if (misc_pwr & (1 << 8)) {
2583 			pr_debug("Bit 8 in the MISC_PWR_MGMT MSR set\n");
2584 			return true;
2585 		}
2586 	}
2587 
2588 	idx = acpi_match_platform_list(plat_info);
2589 	if (idx < 0)
2590 		return false;
2591 
2592 	switch (plat_info[idx].data) {
2593 	case PSS:
2594 		if (!intel_pstate_no_acpi_pss())
2595 			return false;
2596 
2597 		return intel_pstate_no_acpi_pcch();
2598 	case PPC:
2599 		return intel_pstate_has_acpi_ppc() && !force_load;
2600 	}
2601 
2602 	return false;
2603 }
2604 
2605 static void intel_pstate_request_control_from_smm(void)
2606 {
2607 	/*
2608 	 * It may be unsafe to request P-states control from SMM if _PPC support
2609 	 * has not been enabled.
2610 	 */
2611 	if (acpi_ppc)
2612 		acpi_processor_pstate_control();
2613 }
2614 #else /* CONFIG_ACPI not enabled */
2615 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2616 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2617 static inline void intel_pstate_request_control_from_smm(void) {}
2618 #endif /* CONFIG_ACPI */
2619 
2620 #define INTEL_PSTATE_HWP_BROADWELL	0x01
2621 
2622 #define ICPU_HWP(model, hwp_mode) \
2623 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode }
2624 
2625 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2626 	ICPU_HWP(INTEL_FAM6_BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL),
2627 	ICPU_HWP(INTEL_FAM6_BROADWELL_XEON_D, INTEL_PSTATE_HWP_BROADWELL),
2628 	ICPU_HWP(X86_MODEL_ANY, 0),
2629 	{}
2630 };
2631 
2632 static int __init intel_pstate_init(void)
2633 {
2634 	const struct x86_cpu_id *id;
2635 	int rc;
2636 
2637 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2638 		return -ENODEV;
2639 
2640 	if (no_load)
2641 		return -ENODEV;
2642 
2643 	id = x86_match_cpu(hwp_support_ids);
2644 	if (id) {
2645 		copy_cpu_funcs(&core_funcs);
2646 		if (!no_hwp) {
2647 			hwp_active++;
2648 			hwp_mode_bdw = id->driver_data;
2649 			intel_pstate.attr = hwp_cpufreq_attrs;
2650 			goto hwp_cpu_matched;
2651 		}
2652 	} else {
2653 		id = x86_match_cpu(intel_pstate_cpu_ids);
2654 		if (!id) {
2655 			pr_info("CPU model not supported\n");
2656 			return -ENODEV;
2657 		}
2658 
2659 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2660 	}
2661 
2662 	if (intel_pstate_msrs_not_valid()) {
2663 		pr_info("Invalid MSRs\n");
2664 		return -ENODEV;
2665 	}
2666 
2667 hwp_cpu_matched:
2668 	/*
2669 	 * The Intel pstate driver will be ignored if the platform
2670 	 * firmware has its own power management modes.
2671 	 */
2672 	if (intel_pstate_platform_pwr_mgmt_exists()) {
2673 		pr_info("P-states controlled by the platform\n");
2674 		return -ENODEV;
2675 	}
2676 
2677 	if (!hwp_active && hwp_only)
2678 		return -ENOTSUPP;
2679 
2680 	pr_info("Intel P-state driver initializing\n");
2681 
2682 	all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2683 	if (!all_cpu_data)
2684 		return -ENOMEM;
2685 
2686 	intel_pstate_request_control_from_smm();
2687 
2688 	intel_pstate_sysfs_expose_params();
2689 
2690 	mutex_lock(&intel_pstate_driver_lock);
2691 	rc = intel_pstate_register_driver(default_driver);
2692 	mutex_unlock(&intel_pstate_driver_lock);
2693 	if (rc)
2694 		return rc;
2695 
2696 	if (hwp_active)
2697 		pr_info("HWP enabled\n");
2698 
2699 	return 0;
2700 }
2701 device_initcall(intel_pstate_init);
2702 
2703 static int __init intel_pstate_setup(char *str)
2704 {
2705 	if (!str)
2706 		return -EINVAL;
2707 
2708 	if (!strcmp(str, "disable")) {
2709 		no_load = 1;
2710 	} else if (!strcmp(str, "passive")) {
2711 		pr_info("Passive mode enabled\n");
2712 		default_driver = &intel_cpufreq;
2713 		no_hwp = 1;
2714 	}
2715 	if (!strcmp(str, "no_hwp")) {
2716 		pr_info("HWP disabled\n");
2717 		no_hwp = 1;
2718 	}
2719 	if (!strcmp(str, "force"))
2720 		force_load = 1;
2721 	if (!strcmp(str, "hwp_only"))
2722 		hwp_only = 1;
2723 	if (!strcmp(str, "per_cpu_perf_limits"))
2724 		per_cpu_limits = true;
2725 
2726 #ifdef CONFIG_ACPI
2727 	if (!strcmp(str, "support_acpi_ppc"))
2728 		acpi_ppc = true;
2729 #endif
2730 
2731 	return 0;
2732 }
2733 early_param("intel_pstate", intel_pstate_setup);
2734 
2735 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2736 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2737 MODULE_LICENSE("GPL");
2738